Top Stories
EU Council Revives “Chat Control 1.0” via Written Procedure, Schedules Parliament Vote Last Day Before Summer Recess
On July 2, 2026, the Council of Ministers adopted a position for a “new” regulation substantively identical to the expired April 3 “Chat Control 1.0” temporary regime, using a written procedure so an act that lapsed in April could be technically “renewed” on paper. The Parliament second reading is now scheduled for the last day before summer recess, when assembling the absolute blocking majority needed to reject it becomes a logistical problem in its own right.
The scope is broad: voluntary monitoring of “internet-based, number-independent communication services such as messenger apps, webmail, and VoIP telephony” using “AI and hash matching for known abuse material or grooming patterns.” Heise notes content and traffic data must be “irrevocably deleted,” with a floor of “no later than twelve months after detection,” and quotes the Council position that scans “will be limited to the absolutely necessary extent” and that “no general, indiscriminate surveillance will take place.” The procedural mechanism matters more than any new substance: an E-Privacy Directive exception that expired in April has effectively been reopened as part of a different legislative act, and Heise frames the move as “closing a looming legal loophole” rather than a fresh debate. It pairs with last week’s EFF and allies v. X filing on the FTC consent decree (published July 2, 2026) into a single “AI is being deployed against speech and identity faster than oversight can keep up” line for the analysis column.
Amazon Will Stop Accepting New Mechanical Turk Customers on July 30
Amazon announced on July 5, 2026 that Mechanical Turk “will close to new customers” on July 30, 2026, while existing customers keep their access. AWS’s statement, quoted by TechCrunch: “Existing customers can continue to use the service as normal. AWS continues to invest in security and availability improvements for Mechanical Turk, but we do not plan to introduce new features.”
Turk is the canonical “human-in-the-loop” training-data primitive for AI, and the timing lines up with the broader collapse of paid human annotation inside the major US labs. TechCrunch describes the service as “very much on life support.” Read alongside the Forbes piece below and Jack Dorsey’s Block announcement from February 26, the practical signal is that the requester pipeline for paid human data work is being wound down quietly while narratives about AI “augmenting” or “freeing up” human workers continue in earnings calls.
Forbes: “AI Costs More Than the People It Replaced”
Jemma Green at Forbes argues the operational cost of running frontier AI in production often exceeds the loaded salary of the human role it nominally replaces, and that “tokenmaxxing” - management incentives that reward raw AI usage over productivity - is making the gap worse. She cites inference cost (OpenAI loses about $2 for every $1 of revenue earned, with ~$14B projected losses this year and ~$44B cumulative losses projected before 2029 profit), integration cost (a 30-50% bill increase expected when subsidized AI pricing normalizes; one company hit a $500M Claude bill in a single month), and supervision cost (Jensen Huang’s reported $2B annual Nvidia internal token budget, with Huang’s framing that “$500K engineers should consume $250K in tokens annually”).
The error-correction data point is the one that lands hardest: Green reports that “code churn” - lines deleted versus added - increased 800%+ under high AI adoption, and frames it as a measurable “feature” of generated code that needs human review. She also cites Sequoia’s David Cahn’s $600B annual revenue figure needed to justify current AI infrastructure spending, and a Gartner forecast of $207B in AI agent software spending in 2026 (up 139%). The Forbes piece is the cleanest mainstream-press counterweight to the “AI cuts costs” line that has been driving the Block 50% layoff and Salesforce 1,000-person round since Q1 - see our Block coverage at the time and Salesforce coverage for the longer read.
Meta Publishes Its AI Storage Blueprint at Scale
Meta engineers Sidharth Bajaj and Venkatraghavan Srinivasan detailed on July 1, 2026, how Meta’s BLOB-storage architecture evolved to keep GPU utilization high and research velocity fast for AI workloads. The new stack flattens Meta’s old stateful metadata layers (namelayer, volumeslayer, containerlayer) onto a horizontally scalable block layer called Tectonic, with ZippyDB backing a unified metadata schema and Meta’s Owl subsystem reused for distributed GPU-host caching. Reported numbers include an “average cache hit rate of 80% on the distributed data cache” and a read-plan metadata cache with “1-2 ms access.”
The notable admission is the framing. Meta writes “if AI is the brain, storage is the memory” - meaning the company is now treating its storage layer as a first-class part of the AI training stack, not as a sidecar to GPU boxes. The post describes “hundreds of exabyte-scale storage clusters” powering Meta’s products, and says the redesign was driven by the goals of being able to “ingest data once and access data anywhere” and “iterate in minutes and not hours.” It is the first time Meta has published concrete numbers for the storage half of its AI infra stack, and useful reference material for anyone tracking where the AI capex dollar is actually landing beyond GPU SKUs.
Hugging Face Ships “Kernels: Major Updates” Across the Hub
The Hugging Face blog rolled up a major set of changes on July 6, 2026, to the Hub’s kernel packaging, distribution, and consumption pipeline. The new releases introduce a dedicated “kernel” repository type on the Hub (e.g. kernels-community/flash-attn3), kernel signing using “Sigstore’s cosign to sign using ephemeral private keys,” and a split CLI architecture separating kernels (load) from kernel-builder (build).
The local-ai signal is that HF Jobs and the Hub are starting to ship compiled kernels rather than just model weights, which closes a long-standing gap with Ollama and llama.cpp’s “ship the runtime with the model” model. Framework support widened as well: “Apache TVM FFI is the first framework to be supported besides Torch,” and kernels now link libstdc++ dynamically. The post also lays the foundation for agentic kernel development - an agent-optimized CLI and backend-specific skills - which makes the Hub look more like a kernel-runtime registry than a model registry over time.
Study: AI-Generated Fiction Strips the Mystery Out of Storytelling
Researchers at UNC Chapel Hill built CASPER, an automated framework that scores character portrayal across eight literary dimensions, and found that AI-generated fiction “plays it safe” - wrapping storylines neatly, leaning on predictable archetypes, and avoiding the ambiguity readers associate with “lingering” stories. The lead author is graduate student Anneliese Brei, with undergraduate co-author Nicholas Sanaie and senior author Snigdha Chaturvedi (associate professor); all are at UNC-Chapel Hill. The paper notes that scaling doesn’t fix the deficit: larger flagship LLMs produced “characters just as flat as smaller models.”
CASPER is explicitly positioned as a benchmark for “evaluating whether future models advance narrative depth or merely become more grammatically fluent.” For analysis purposes, this is the cleanest available data point behind the long-running “AI writing feels flat” intuition, and it sits naturally next to last week’s GPT-4 Turbo impersonation study at 404 Media (published July 1, 2026) - that one found AI wins at faking authenticity in real political discourse, while this one finds AI loses at producing the ambiguity that makes written fiction feel human. The two together describe a clean dividing line between fluency and depth.
Quick Hits
- Mechanical Turk requester pipeline winds down quietly: Amazon’s July 5 announcement describes the service as one the company “do[es] not plan to introduce new features” for; new-customer sign-ups close July 30, 2026.
Worth Watching
The EU’s procedural sprint is the real story. Fast-tracking an expired temporary regulation by routing a substantively identical “new” act through a written procedure, then scheduling a Parliament vote on the last day before summer recess, is the kind of single-move decision that defines a regulatory era. Watch the Parliament second reading for whether the absolute blocking majority of MEPs shows up; the result sets the precedent for every other “renewed” scanning regime the Council wants to extend.
The labor-cost inversion is starting to be measurable. Mechanical Turk closing to new customers, the Forbes “AI costs more than the people it replaced” piece, and the UNC CASPER “generated fiction has no depth” study all sit in the same window. The clean analysis column for the week pairs the Forbes piece with a real P&L breakdown of a named company that did the human-to-AI swap, including inference, integration, supervision, and error correction. Until that column exists, “AI cuts costs” remains a slogan rather than a number.
Storage is the new compute. Meta’s blueprint is the largest AI infra-vendor admission to date that storage, not FLOPS, is the binding constraint on AI training. Combined with the earlier Meta-AMD 100B chip deal, the practical read is that AI capex is migrating sideways into bandwidth, power, and environmental permits - bottlenecks the GPU boom never had to clear.